Computer vision · Visual quality control
Visible defects,flagged in time.
Industrial visual inspection with AI is automated, image-based quality control: a camera above the line photographs every part, and a model trained on your examples of good and defective products marks what is worth checking. The operator confirms every rejection, and the images remain as evidence for quality analysis.
The eye gets tired, and sampling leaves gaps
On a fast line, an operator cannot look at every part with the same attention for the whole shift. Sample-based checks only find defects in the sample, and when a problem appears, there are rarely images showing since when and how often it happened.
- Scratches, cracks, stains or wrong labels go unnoticed between two checks
- Rejection criteria differ from one operator to another and from one shift to another
- Complaints arrive after the batch has shipped, with no evidence of the parts' condition at dispatch
- There are no clear statistics on defect types to help find the cause
Who it is for
- Factories with production lines where visual checks are done manually or by sampling
- Quality departments that need visual evidence of non-conformities
- Manufacturers of components, packaging, cast or injection-moulded parts with recurring surface defects
- Process engineers who want data on defect types and frequency by shift and batch
What defects visual inspection can flag
- 01
Flags surface defects
Marks visible scratches, cracks, stains or deformations, based on examples labelled by the quality team.
- 02
Checks assembly and labelling
Checks whether a component is missing, whether the label is present and correctly positioned or whether a cap is fitted.
- 03
Shows the suspect area
The area considered defective is highlighted on the image, so the operator can quickly see what triggered the flag.
- 04
Confirmation queue at the station
Flagged parts appear on a screen next to the line, where the operator confirms or cancels the rejection.
- 05
Image archive by batch
Images of flagged parts are kept with the batch, shift and time, for traceability and complaint analysis.
- 06
Defect statistics
The frequency of each defect type by shift and batch is displayed for process engineers.
How it works
What happens to each part on the line
The steps marked in yellow are approved by a person.
- Input
- AI
- Action
- Human in the loop
- Result
- 01
Input
The part enters the frame
A sensor or the line signal triggers the camera at the right moment.
- 02
AI
Visual analysis
The model compares the part with what it learned from good and defective examples.
- 03
Action
Flag at the station
Suspect parts appear with the area marked on the operator's screen.
- 04
Human in the loop
The operator confirms the rejection
A person decides whether the part is rejected, rechecked or accepted.
- 05
Action
Recorded in the log
The decision and the image are saved together with the batch and shift.
- 06
Result
Quality data
The team sees defect types and how they change over time.
What examples, cameras and lighting are needed
- Real examples of good parts and of parts with each type of defect, labelled by quality specialists
- A defect catalogue with clear acceptance and rejection criteria
- An industrial camera chosen for the line speed and defect size, with a lens suited to the working distance
- Controlled, consistent lighting (diffuse, backlit or angled), often more important than the camera
- A trigger signal from the line (sensor, encoder or PLC) and a stable mount, protected from vibration
- An industrial computer or edge device next to the line, for low-latency processing
Possible integrations
- GigE Vision or USB3 Vision industrial cameras
- PLCs via OPC UA
- Modbus TCP
- MES systems
- ERP for batch traceability
- NVIDIA Jetson edge devices
- Industrial computers with GPUs
- Grafana or Power BI dashboards
- SQL databases for the image archive
- GigE Vision or USB3 Vision industrial cameras
- PLCs via OPC UA
- Modbus TCP
- MES systems
- ERP for batch traceability
- NVIDIA Jetson edge devices
- Industrial computers with GPUs
- Grafana or Power BI dashboards
- SQL databases for the image archive
Each integration's compatibility is checked during the analysis phase, for the systems and versions you use.
Control and transparency
What stays with people
- Every rejection, confirmed by the operator before the part is removed from the flow
- Defining and updating the defect catalogue and the acceptance criteria
- Validation by the quality team of the labels used for training
- The decision to stop the line or hold a batch, which stays with the production or quality manager
- Regular sample-based re-inspection of parts accepted by the system, to uncover missed defects
Limits and operating conditions
- Inspection covers defects visible in the image; internal defects or dimensional deviations below the camera's resolution need other inspection methods
- Rare defects, for which there are few examples, are harder for the model to learn
- Glossy, transparent or textured surfaces need carefully tested lighting and positioning
- A change of material, colour or supplier may require new examples and retraining
- The system supports existing quality control procedures and certifications; it does not replace them
How we evaluate inspection: missed defects and false alarms
We agree the indicators together before launch and track them afterwards. The values depend on the process, the data and the volume.
- Missed defects
- Defective parts accepted by the system, identified through manual re-inspection of a sample of passed parts
- False alarms
- The share of flagged parts the operator accepted as good, recorded with each confirmation
- Agreement with inspectors
- A comparison between the model's results and quality inspectors' decisions on the same set of parts
- Keeping up with the line
- Analysis time per part, measured in production, against the time available between two parts
How we implement visual inspection on the line
- 01
We catalogue the defects
We work with the quality team to define defect types, criteria and relevant examples.
- 02
We test the camera and lighting
We run trials at the line, with different angles and lighting, until defects become consistently visible.
- 03
We train on real parts
We build the dataset from production and test the model on parts it has not seen.
- 04
Pilot with the operator in the loop
We run it on a single line, the operator confirms every flag, and we adjust the thresholds based on their decisions.
Use cases
The scenarios below are conceptual examples, designed to illustrate realistic applications. They are not implementations delivered by WEB-AZ.
- Automotive components
Scratches on injection-moulded plastic parts
At the injection press output, the camera photographs every part, and visible scratches are marked for the operator at the station.
Conceptual example - Bottling
Missing or misaligned labels
On the bottling line, the system flags bottles without a label or with a crooked label, and the operator decides whether to remove them.
Conceptual example - Woodworking
Visible defects on panels
Panels are photographed before packaging, and marked knots or cracks go to an inspector, who sets the quality grade.
Conceptual example
Useful guides
Before you decide, read
Frequently asked questions
What is industrial visual inspection with AI?
It is quality control carried out with cameras and a computer vision model: each part is photographed on the line, and the model flags visible defects learned from your examples, such as scratches, cracks, stains or missing labels. The operator confirms rejections, and the images are kept for traceability.
Does the system reject parts on its own?
Not in the configuration we propose. The system flags, and the operator confirms the rejection. Any automatic diversion mechanism is discussed separately, after performance has been measured, and only with the quality team's approval.
What happens when the model is not sure whether a part is defective?
Parts with an uncertain score are treated as suspect and go to the operator. Their decisions are recorded and can be used to improve the model during retraining.
Do we need new cameras?
Usually, yes. Inline inspection needs industrial cameras and dedicated lighting, chosen for the line speed and defect size; surveillance cameras are rarely suitable. We run trials before recommending equipment.
How many defect examples do we need to collect?
It depends on how many defect types there are and how varied they are. For rare defects, we can start with an approach that learns what a good part looks like and flags anything that looks different.
Where are production images stored?
By default, locally, on the equipment next to the line or on a server in the factory. Access and retention are agreed with you, and the images are not used for other purposes without your consent.
Can’t find the answer? Write to us and we’ll reply directly.
Do you have defects that slip past visual checks?
Send us examples of good and defective parts, plus a few details about the line. We will tell you whether the defects can be captured visually and what equipment would be needed.
Related solutions
- Object detectionModels that recognise the objects that matter to your process in images and pass them on as structured data.
- Edge AI for computer visionVideo analysis on devices installed on site: fast response, less network load and images that never leave the building.
- Operational monitoringVisual tracking of process stages and of vehicle and pallet movement in yards and warehouses, so you can see where the flow gets stuck.